Deep Learning For Change Detection In Multitemporal Remote Sensing Images

With the increasing improvement of space observation capability,remote sensing image change detection technology has been an important means for dynamically monitoring the Earth's environmental changes,which has been widely used in urban expansion evolution analysis,agricultural survey,natural disaster evaluation and so on.As the main type of current remote sensing data,multitemporal remote sensing images contain abundant spatial and temporal information,and have the characteristics of volume,high-dimensionality and non-construction.However,how to solve the problems of learning and optimization in multitemporal remote sensing image change detection is still the main challenge we are facing at present.Therefore,the thesis focuses on studying the project of remote sensing image change detection in depth,aiming to develop deep learning-based models and algorithms for the joint interpretation of multitemporal and multi-source spatial-temporal images,thus providing efficient approaches for the extraction and application of remote sensing knowledge.The concrete content includes the following aspects:(1)Synthetic aperture radar(SAR)has the imaging ability in all time and at all weather,which can accurately obtain the ground objects information and is not affected by the weather and illumination conditions,and has been an important data source of remote sensing images.Due to the fact that it employs the principle of coherent imaging for capturing the ground information,however,these acquired images are often contaminated by speckle noise,which greatly increases the difficulty in accurately detecting the land surface changes.Because traditional change detection methods cannot balance the change detail preservation and the speckle noise restriction well,we propose two decision level change detection algorithms based on convolution neural networks.Using different ratio-based operators to generate multiple difference images,based on preclassification and the strategy of sample selection,obtaining training samples with high confidence to construct convolution neural network models and thus achieving the goal of distinguishing changed pixels from unchanged ones.The experimental results demonstrate that both of methods are able to effectively detect changes occurred on the ground as well as to suppress the interference of speckle noise,and thus obtaining accurate detection results.(2)Compared with SAR images,very-high-resolution(VHR)multispectral remote sensing images contain abundant spectral information,which can clearly illustrate the type information of ground objects and is easy to obtain,and therefore has been the important data source for observing the Earth's environmental changes.However,how to exploit the spectral information for monitoring the land use/land-cover changes is a challenging and open problem.To address this problem,we present a novel superpixel-based difference representation learning framework for change detection.It uses the superpixel segmentation technique to segment multitemporal remote sensing images,and takes a single superpixel(i.e.,image-object)as the basic analysis unit,which aims at extracting the change information from spectrum,texture and spatial structure of images,and building deep neural networks to further determine the actual changed regions as well as to eliminate the spurious changes and outliers.The proposed model can achieve superior performance on VHR multispectral data sets obtained by different sensors,and identify the real changed regions accurately.(3)For different types of remote sensing images,we usually need to implement image modeling and analysis,and design suitable algorithms to highlight changed regions.This process is time-consuming and laboring,which great limits their application in real scenarios.Nowadays,SAR images and optical images have been the main data source for monitoring our planet environmental changes,it is necessary to devise a common change detection framework for both types of images above.To this end,we employ a well pretrained deep fully convolutional network model to extract high-level feature representations from raw data automatically,and carry out the uncertainty analysis by combining the spectral information,followed by using the technique of multiscale image segmentation and the strategy of sample selection to obtain valid training samples,and thus training a robust support vector machine(SVM)classifier to fulfill the discrimination of changed regions.The experimental results demonstrate that this method is insensitive to the types of remote sensing images,and it can obtain satisfactory results on different data sets.(4)Actually,most of the acquired remote sensing image-pairs are from the same sensor,we generally called them homogeneous images.However,in some practical emergencies(such as earthquakes,floods,typhoons,etc.),due to the limitations of weather and image acquisition conditions,we cannot get the homogeneous images before and after the event timely.Therefore,how to exploit the image-pair from different sensors for change detection is of great significance and practical value.To solve it,we design a novel change detection method based on logarithmic transformation feature learning,which applies logarithmic transformation on the SAR image,aiming to confer it with similar statistical distribution characteristics as the optical image.On this basis,we utilize the joint feature learning technique for feature extraction and comparison in an unsupervised manner,desiring to improve the identification of changed regions by using a deep neural network.The experimental results show that this method is able to effectively detect changed regions on data sets achieved from different sensors,which further validates its feasibility.(5)Conventional change detection methods only detect changed regions,neglecting the deep analysis of them.In the field remote sensing data analysis,it is difficult to discriminate different types of changes effectively.To address this problem,we propose an iterative feature mapping network framework for detecting multiple changes.It can extract high-level feature representations from raw data based on feature extraction techniques in an unsupervised way,and apply the random initialization on the obtained feature sets for iteratively feature mapping learning so as to yield more discriminative features.Through the hierarchical clustering analysis technique,we can effectively distinguish the type information of changed regions as well as the relationship of different types of changes.The experimental results over homogenous and heterogeneous data sets validate the superiority of the proposed method,which contributes to promoting the subsequent image interpretation and analysis of changed regions.